baoyu-url-to-markdown: a Claude skill that saves any web page as Markdown

• Asks you questions with the agent's own question tool

  When the skill needs an answer from you, it uses a built-in question tool such as AskUserQuestion. If the agent has none, it prints a numbered list and waits for your reply.

  User input rule

• Finds Bun and installs the script's packages on first run

  Claude finds the skill folder, checks that bun is installed (or suggests installing it), and runs bun install on the scripts folder if node_modules is missing.

  Runs code · CLI setup

• Reads your saved settings from one of three folders

  It checks .baoyu-skills/baoyu-url-to-markdown/EXTEND.md in the project, then the XDG config folder, then ~/.baoyu-skills. The first one found wins. Settings are download_media (ask, 1 or 0) and default_output_dir.

  Files · Preferences (EXTEND.md)

• Blocks conversion until you answer three questions

  If no EXTEND.md exists, Claude must ask about media downloads, the output folder and where to save preferences. It then writes EXTEND.md. It may not create one with silent defaults.

  Write · First-time setup

• Headless Chrome fetch, Markdown printed to the screen

  Runs baoyu-fetch with the URL. Chrome runs in the background and the Markdown goes to stdout.

  Runs code · Default capture

• Writes the Markdown to a path Claude builds

  Adds --output. Claude builds the path itself, by default ./url-to-markdown/{domain}/{slug}/{slug}.md, and adds a timestamp if the file already exists.

  Runs code · Save to a file

• Saves images and videos next to the Markdown file

  Adds --download-media, which needs --output. Images go to imgs/ and videos to videos/, and the Markdown links are rewritten to local paths. With the default setting, Claude asks you first.

  Runs code · Download media

• Opens a visible Chrome window and waits for you

  --wait-for interaction detects login, CAPTCHA and Cloudflare gates, waits for you to clear them, then continues. Waits up to 10 minutes by default.

  Runs code · Wait for login or CAPTCHA

• Opens Chrome and waits until you press Enter

  --wait-for force is for pages that need scrolling, paywall steps or other manual browsing. You press Enter when the page is ready.

  Runs code · Manual control

• Returns structured data instead of Markdown

  --format json (or --json) adds fields such as adapter, status, login state, the document, media and the Markdown text.

  Runs code · JSON output

• Picks x, youtube, hn or generic by hand

  Adapters are chosen from the URL. --adapter overrides the choice.

  Runs code · Force an adapter

• Timeouts, Chrome path, profile folder and debug output

  --timeout (30000 ms), --interaction-timeout (600000 ms), --cdp-url to reuse a Chrome, --browser-path, --chrome-profile-dir, --media-dir, --headless and --debug-dir, which writes document.json, markdown.md, page.html and network.json.

  Runs code · Other options

• Tweets, threads and X Articles, with login detection

  For x.com and twitter.com. On pages that need a sign-in it returns needs_interaction, and you rerun with --wait-for interaction.

  Network · X / Twitter adapter

• Transcript with [MM:SS] stamps and chapter headings

  For youtube.com and youtu.be. It needs an exposed caption track. With captions off or restricted playback, you may get only the description.

  Network · YouTube adapter

• Story details and nested comment threads

  For news.ycombinator.com. It returns title, URL, author, score, comment count and the nested replies, and marks deleted or dead items.

  Network · Hacker News adapter

• Any other URL, with Defuddle and a fallback extractor

  Uses Defuddle, falls back to Readability, scrolls the page and waits for the network to go idle.

  Network · Generic adapter

• Claude checks every headless result before trusting it

  Headless Chrome can return login walls or empty shells with exit code 0. Claude checks the title, the body and failure signs, and retries with a wait mode if the page looks wrong.

  Quality gate

• One variable, plus fixes for common failures

  BAOYU_CHROME_PROFILE_DIR sets the Chrome profile folder. Chrome not found: use --browser-path. Timeout: raise --timeout. Login or CAPTCHA: use --wait-for interaction. Debugging: use --debug-dir.

  Files · Environment and troubleshooting

Watch out for

• It saves your X session cookies in its Chrome profile

  After a run on an X or Twitter URL, the code copies the auth_token and ct0 cookies into x-session-cookies.json in the profile folder, and restores them on later runs. Keep that folder private and do not sync it. github.com

• Some page URLs may go to the defuddle.md service

  When local extraction fails or looks incomplete, the code requests https://defuddle.md/ followed by the page URL, so that service sees the URL. Avoid it for private addresses. github.com

• It installs packages with bun on first run

  SKILL.md has Claude run bun install on the scripts folder. package.json lists 10 third-party packages, including jsdom, defuddle and chrome-launcher. github.com

• It launches Chrome and runs a bundled script

  baoyu-fetch is a shell wrapper that runs bun on lib/cli.ts. The code starts Chrome with a debugging port on 127.0.0.1 and closes it afterwards. On macOS it can bring a browser window to the front with osascript. If a launch fails and the profile folder has stale Chrome lock files (SingletonLock, SingletonSocket, SingletonCookie, chrome.pid), it deletes them and retries once. github.com